Jupyter Notebook Tutorial for zero shot text classification

[bhaskar2443053]

This issue is part of #691
This PR fixes #734

Description of changes

This tutorial is an introduction to the Zero-Shot Learning technique for text classification.

Possible influences of this PR.

removed text_classification_pipeline.ipynb , as its updated with a new tutorial

Test Conducted

Describe what test cases are included for the PR.

[codecov]

Codecov Report

Merging #799 (280b574) into master (6ddbcab) will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master     #799   +/-   ##
=======================================
  Coverage   80.50%   80.50%           
=======================================
  Files         252      252           
  Lines       19363    19363           
=======================================
  Hits        15588    15588           
  Misses       3775     3775           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 6ddbcab...280b574. Read the comment docs.

[hunterhector]

Since the tutorial is quite large let me comment them here instead of inline:

1. The rendering format is a bit off, you can check it here: https://asyml-forte--799.org.readthedocs.build/en/799/notebook_tutorial/zero_shot_text_classification.html, notice that the bullet list format is not working
2. The explanation of the pipeline is a bit too rigid. You introduce the whole configuration list, but that's a bit too lengthy and we should probably talk about the code, not what the configuration parameters are.
3. There are grammar errors from time to time, like from hugging face model, it should be from the HuggingFace model. Instead of Zero shot is a powerful tool, it should be Zero Shot model is a powerful tool
4. The cell output is really long, I don't think we want to include all those.

Overall, there is an improvement to the writing, but it still does not feel like a tutorial to me.

[hunterhector]

The sample 10k is a bit too large for the repo (4MB), which would increase the git history by quite a bit. Let's consider hosting it at a different place?

[hepengfe]

Or just keep a few samples (~20) is good enough for the tutorial.

[bhaskar2443053]

removed the 10k sample and added a 20 rows sample.

[hepengfe]

Could you check grammar errors? I have seen some apparent grammar errors.

[bhaskar2443053]

I fixed using Grammarly.

[hepengfe]

for package name or suffix, maybe you can mention them by using double backquotes like sklearn and csv rather than add them into spelling wordlist.

[bhaskar2443053]

used double backquotes to fix spellings not in wordlist

[bhaskar2443053]

I am unable to fix the bullet points. I tried '- ' , '* ' and '1. ' , but none of them are rendering correctly in docs build.
any suggestions?

[hepengfe]

I am unable to fix the bullet points. I tried '- ' , '* ' and '1. ' , but none of them are rendering correctly in docs build.
any suggestions?

According to https://buildmedia.readthedocs.org/media/pdf/nbsphinx/latest/nbsphinx.pdf,

In case that your Bulletpoints do render in the notebook and do not render with nbsphinx, please add one blank line before the bulletpoints.

You can google the notebook issue with keyword nbsphinx which is the package we are using to convert the notebook to html.

[hepengfe]

I have a few suggestions.

1. You could remove square brackets around the mentioned issues in the PR description.
2. Some descriptions such as the one below is inaccurate or don't relate very well to forte. I would suggest refer to some official definition and rephrase it to show how forte could handle data and organize nlp components for this particular task.

Text classification is the process of grouping text data into various categories based on some algorithm. These algorithms are at the heart of a variety of software systems that process text data at scale

3. maybe backtick the pipeline component in the markdown cell such as NLTKSentenceSegmenter( )

In general, I would suggest review both the notebook and the rendered html page after you are done with editing.

[bhaskar2443053]

I am unable to fix the bullet points. I tried '- ' , '* ' and '1. ' , but none of them are rendering correctly in docs build.
any suggestions?

According to https://buildmedia.readthedocs.org/media/pdf/nbsphinx/latest/nbsphinx.pdf,

In case that your Bulletpoints do render in the notebook and do not render with nbsphinx, please add one blank line before the bulletpoints.

You can google the notebook issue with keyword nbsphinx which is the package we are using to convert the notebook to html.

adding a blank line above bullets worked.

[bhaskar2443053]

3. backtick

Mostly fixed all the above comments.

[hepengfe]

Some minor issues. After these are fixed I think it's ready to merge.

These algorithms are at the heart of various software systems that process text data at scale.

I think we probably don't want to keep this sentence as it doesn't relate very well to either the context or forte.

body is the complete description of the review.

body -> Body

where the model tries to classify without any fine-tuning

without any fine-tuning on a single data.

let’s also check the class balance of the data by using groupby and count functions in pandas.

let -> Let

The first entry to our pipeline will be the data reading step.

Rather than entry, we want to refer them as component or pipeline component. As entry is referring to something else in forte.

Accuracy of 95% on unseen data is an exceptional performance from the Hugging-Face model. Zero-Shot learning is a powerful tool for low volume of label data.

I think we don't need to backtick HuggingFace and zero-shot learning as they are already known names. By backtick package names, I mean suppose transformers as a python package name shown in markdown cell, we definitely need to backtick it.

[bhaskar2443053]

Some minor issues. After these are fixed I think it's ready to merge.

These algorithms are at the heart of various software systems that process text data at scale.

I think we probably don't want to keep this sentence as it doesn't relate very well to either the context or forte.

body is the complete description of the review.

body -> Body

where the model tries to classify without any fine-tuning

without any fine-tuning on a single data.

let’s also check the class balance of the data by using groupby and count functions in pandas.

let -> Let

The first entry to our pipeline will be the data reading step.

Rather than entry, we want to refer them as component or pipeline component. As entry is referring to something else in forte.

Accuracy of 95% on unseen data is an exceptional performance from the Hugging-Face model. Zero-Shot learning is a powerful tool for low volume of label data.

I think we don't need to backtick HuggingFace and zero-shot learning as they are already known names. By backtick package names, I mean suppose transformers as a python package name shown in markdown cell, we definitely need to backtick it.

modified according to the above feedbacks.